Claude Code: after two failed fixes, run /clear

The tutorial frames the problem as one of resource management: the context window holds everything in the conversation, including failed code and the corrections pointing it out. Because Claude reads all of it when composing the next reply, those failed approaches can become reference material that steers the next attempt wrong. The documentation's own language is stark — a clean session plus a good prompt almost always beats a long correction-heavy session. The tutorial's recommended fix is not just /clear but a rewritten prompt that adds three things learned from the failures: the goal, the place to make the change, and the approach to avoid. It gives the example of a login redirect in Next.js, where Claude kept adding router.push inside useEffect; the restarted prompt specifies middleware.ts and rules out useEffect transitions. Beyond /clear, the piece maps out the other commands: /compact summarizes while letting you say what to keep, /btw asks a side question without adding to the main history, and CLAUDE.md holds rules for what summaries must preserve. The author notes using /rewind, /compact with instructions, and /clear to split a problem into smaller sessions.